I had this same doubt when I was studying the Thompson's construction. As I see I am not the only one, I will try to solve the mystery.
Consider the regular expression: $$(a|b)|(c|d)$$
With Thompson's construction we generate first:
Now let's use the noxwox's construction that you suggested. First we have:
Can you see the difference? I had a first suspicion watching these two results. Then I reviewed the Thompson's constructions and I noticed something interesting. We can see the NFAs as directed graphs and the Thompson's construction guarantees a graph whose nodes have at most two successors.
So here starts my conclusion: Generating the data structure to store a NFA obtained by Thompson's construction is very easy because it is a set of nodes with two pointers each. If we use nowox's construction we don't know a priori the numbers of successors of each node and we have to change dynamically the amount of memory reserved for each node or be inefficient in the memory management. From this point of view the Thompson's construction algorithm guarantees a graph that is easy and fast to generate in a computer and I think that the additional computational cost of having more states than the NFA generated by nowox's construction is overshadowed by the backtracking mechanic of the NFAs.